Systems and open-source engineer focused on columnar query processing, distributed ML training determinism, and AI agent runtime verification.
I spend most of my time writing query execution kernels in the Apache Arrow and DataFusion ecosystems, building spatial routing engines, and contributing fixes to open-source data infrastructure.
A live snapshot of contributions across data systems, ML infrastructure, and runtime verification:
- Merged Upstream (34 PRs): Merged code in Apache DataFusion, DataFusion-Spark, Hugging Face Diffusers, DevTrack, AgentWatch, and the NSoC ecosystem.
- Open Upstream (21 PRs): Active PRs under review across Apache Arrow-rs, Apache DataFusion, PyTorch Lightning, SafeTensors, and Diffusers.
- Repositories Contributed (12+): Contributions spanning the Apache Software Foundation, Hugging Face, PyTorch / Lightning AI, and AI agent platforms.
Columnar query execution, arithmetic kernels, and engine compatibility across the Apache ecosystem.
- Apache DataFusion: Merged fix for benchmark row-count propagation in scalar GCD Criterion benchmarks (#26465); submitted cumulative
Utf8array offset overflow protection inrepeat(#27643) and partition metadata handling fixes (#27561). - DataFusion-Spark: Merged compatibility fixes matching Spark 4 string charset encoding (#116), datetime interval overflow handling (#118, #119), and active port overflow validation in
parse_url(#390). - Apache Arrow (Rust): Submitted decimal safe-cast overflow handling under extreme scale expansion in
arrow-cast(#12389) and strict duplicate key rejection in Parquet Variant parsing (#12135).
Autonomous satellite radar (SAR) oil spill segmentation and hydrodynamic drift hindcasting platform.
- Ingestion pipeline for Copernicus Sentinel-1 SAR radar imagery, applying 7x7 Lee speckle filtering and adaptive local Gaussian thresholding to isolate oil slicks from sea clutter.
- Dual lightweight ONNX Runtime CPU models (
SpillUNetandSceneClassifier, ~105 KB total payload) for zero-GPU edge inference. - 2D Euler hydrodynamic drift model simulating slick advection and dispersion driven by surface wind shear and ocean current vectors (Open-Meteo CMEMS/MET Norway).
- Full-stack geospatial architecture combining FastAPI, React, Leaflet GIS, and SQLite/PostgreSQL caching.
Terrain- and weather-aware dynamic multi-modal freight routing engine and forward supply chain decision support.
- Dynamic multi-modal graph routing engine (A* search) factoring in Digital Elevation Model (DEM) slope penalties and physical speed floors across difficult terrain.
- 7-factor dispatch priority queue balancing cargo perishability, medical criticality, contract SLAs, transport cost, and route reliability.
- PostgreSQL and PostGIS spatial queries (
ST_DWithin,ST_LineInterpolatePoint) for real-time waypoint tracking, backed by a 280-test regression suite.
Quantitative time-series forecasting terminal and backtester built in native Scilab with companion Python models.
- Pure Scilab numerical engine implementing Ordinary Least Squares (OLS) polynomial trend fitting with exact 95% prediction intervals, Autoregressive AR(p) modeling, and Holt double exponential smoothing.
- Chronological multi-fold walk-forward validation across market regimes evaluated against naive persistence baselines.
- Companion work in Python (
stock_prediction_python) evaluates LSTM sequence architectures alongside statistical baselines.
Decentralized measurement, reporting, and verification (MRV) registry for blue carbon restoration credits.
- Solidity 0.8.20 smart contracts enforcing cryptographic verification role separation so field verifiers cannot issue credits for their own assessments.
- Immutable, content-addressed storage integration via IPFS (Kubo RPC & Pinata) for remote-sensing evidence and field survey proofs.
- GitHub: @Subodh-17
- LinkedIn: Subodh
- Instagram: @subodh17_
- Email: subodh171011@gmail.com

